Earticle

최근 도로상의 비디오 영상에서 다양한 저차원 정보를 바탕으로 자동 차량 흐름 파악과 사고 탐지에 관해 관심이 높아지고 있다. 본 논문에서는 CCTV 동영상 교통관리시스템을 이용하여 자동으로 차량의 흐름을 파악하고 이를 이용하여 교통체증과 더 나아가 사고탐지를 위한 알고리즘과 그 응용에 대해 연구하였다. 이를 위하여 우선 움직임 객체를 시공간 관계 모델링을 통해 차량의 실제 궤적과 매핑하고 이를 통해 차량의 흐름 파악에 사용하였다. 또한 시공간 관계 모델링과 차량의 실제 궤적간 매핑을 위해 TSR (Tangent Space Representation) 알고리즘을 사용하였다. 또한, 객체의 움직임 추출을 위해 먼저 차(Differece)영상을 이용하여 움직임 객체 추출을 하였고, 이를 통하여 객체의 움직임 트래킹(Tracking)을 하고 각 객체에 번호를 부여하여 동시에 여러 객체를 인식시키고 이를 저장하였다. 이를 통하여 의미적 움직임 객체를 인식하고, 차량 흐름을 파악하는 어플리케이션을 구현하였다. 본 연구를 통해 기본적인 움직임 객체에 대한 의미적인 결과를 얻을 수 있었고, 나아가 CCTV를 이용한 자동 사고 탐지에 관한 연구로 확장시킬 수 있을 것으로 기대된다.

Recently, there are interested in the automatic traffic flowing and accident detection using various low level information from video in the road. In this paper, the automatic traffic flowing and algorithm, and application of traffic accident detection using traffic management systems are studied. To achieve these purposes, the spatio-temporal relation models using topological and directional relations have been made, then a matching of the proposed models with the directional motion verbs proposed by Levin’s verbs of inherently directed motion is applied. Finally, the synonym and antonym are inserted by using WordNet. For the similarity measuring between proposed modeling and trajectory of moving object in the video, the objects are extracted, and then compared with the trajectories of moving objects by the proposed modeling. Because of the different features with each proposed modeling, the rules that have been generated will be applied to the similarity measurement by TSR (Tangent Space Representation). Through this research, we can extend our results to the automatic accident detection of vehicle using CCTV.

To deal with the problems of occlusion, pose variations and illumination changes in the object tracking system, a regression model weighted multi-templates mean-shift (MS) algorithm is proposed in this paper. Target templates and occlusion templates are extracted to compose a multi-templates set. Then, the MS algorithm is applied to the multi-templates set for obtaining the candidate areas. Moreover, a regression model is trained to estimate the Bhattacharyya coefficients between the templates and candidate areas. Finally, the geometric center of the tracked areas is considered as the object's position. The proposed algorithm is evaluated on several classical videos. The experimental results show that the regression model weighted multi-templates MS algorithm can track an object accurately in terms of occlusion, illumination changes and pose variations.

Visual object tracking is a fundamental problem in the field of computer vision, as it needs a proper model to account for drastic appearance changes that are caused by shape, textural, and illumination variations. In this paper, we propose a feature-based visual-object-tracking method with a sparse representation. Generally, most appearance-based models use the gray-scale pixel values of the input image, but this might be insufficient for a description of the target object under a variety of conditions. To obtain the proper information regarding the target object, the following combination of features has been exploited as a corresponding representation: First, the features of the target templates are extracted by using the HOG (histogram of gradient) and LBPs (local binary patterns); secondly, a feature-based sparsity is attained by solving the minimization problems, whereby the target object is represented by the selection of the minimum reconstruction error. The strengths of both features are exploited to enhance the overall performance of the tracker; furthermore, the proposed method is integrated with the particle-filter framework and achieves a promising result in terms of challenging tracking videos.

Mobility arises naturally in the Internet of Things networks, since the location of mobile objects, e.g., mobile agents, mobile software, mobile things, or users with wireless hardware, changes as they move. Tracking their current location is essential to mobile computing. To overcome the scalability problem, hierarchical architectures of location databases have been proposed. When location updates and lookups for mobile objects are localized, these architectures become effective. However, the network signaling costs and the execution number of database operations increase particularly when the scale of the architectures and the numbers of databases becomes large to accommodate a great number of objects. This disadvantage can be alleviated by a location caching scheme which exploits the spatial and temporal locality in location lookup. In this paper, we propose a hierarchical location caching scheme, which acclimates the existing location caching scheme to a hierarchical architecture of location databases. The performance analysis indicates that the adjustment of such thresholds has an impact on cost reduction in the proposed scheme.

Three-dimensional (3D) object tracking is crucial in computer vision applications, particularly in autonomous driving, robotics, and surveillance. Despite advancements, effectively utilizing multimodal data to improve multi-object detection and tracking (MODT) remains challenging. This study introduces ACMODT, an affinity computation-based multi-object detection and tracking framework that integrates camera (2D) and LiDAR (3D) data for enhanced MODT performance in autonomous driving. This approach leverages EPNet as a backbone, utilizing 2D?3D feature fusion for accurate proposal generation. A deep neural network (DNN) extracts robust appearance and geometric features, while an improved affinity computation module combines Refined Boost Correlation Features (RBCF) and 3D-Extended Geometric IoU (3D-XGIoU) for precise object association. Motion prediction is refined using a Kalman filter (KF), and Gaussian Mixture Model (GMM)-based data association ensures consistent tracking. Experiments on the KITTI car tracking benchmark for quantitative analysis and the RADIATE dataset for visualization demonstrate that our method achieves superior tracking accuracy and precision compared to state-of-the-art multi-object tracking (MOT) approaches, proving its effectiveness for real-time object tracking.

This study proposes a mother-slave robot tracking system that identifies the target, predicts its location, and tracks it based on a single image. The proposed system utilizes a Convolutional Neural Network (CNN) for object detection, to identify the target robot. The distance and angle between the robots are then calculated through linear regression analysis, which offers a more efficient and cost-effective solution than traditional methods. The performance of the system was evaluated, resulting in an accuracy of 99.59% for object detection, and an average distance error of 2.04% for the estimated location.

Multi-object tracking techniques are receiving increasing attention due to the growing demand of autonomous driving. Recently, the performance of multi-object tracking has been improved significantly thank to deep learning technique. Most of multi-object tracking methods based on deep learning, however, are highly prone to frequent tracking losses and track-ID switching in case of limited viewpoint and occluded objects. To alleviate this problem, we propose a multi-camera Collaborate Multi Object Tracking (CMOT) method which performs online association of multiple tracked vehicles from stereo vision camera. CMOT not only provides global tracking IDs between multiple cameras but also helps reduce the problem of ID switching compared with the conventional multi-object tracking based on single camera. It can, therefore, improve the overall performance of multi-vehicle tracking compared to each individual camera. We demonstrate the multi-object tracking performance of the proposed method using stereo images of the KITTI dataset.

In autonomous driving, environmental perception and decision-making are technologies that acquire signals based on various sensors and generate information that enables obstacle avoidance, emergency stops, and path planning. Such environmental perception technologies are limited in that they depend on expensive sensors such as LIDAR and RADAR, so research on environmental perception technologies using only cameras is actively being conducted. Studies predicting traffic accidents based on dashcam footage have also been performed as part of such research. This is challenging because only limited forward-looking footage can be acquired, and the surrounding environment is dynamic and changes quickly, making analysis difficult. Existing research has focused on learning the spatialtemporal feature representation to solve these problems. This paper extracts the trajectories of the ego-vehicle and surrounding vehicles using Visual SLAM and Multiple Object Tracking algorithms and uses them as inputs to the graph convolutional neural network(GCN) to learn the spatialtemporal feature representation. In addition, the features learned through the GCN are used as inputs to a bayesian neural network(BNN) to predict the probability of accidents, and its ability to predict accidents in advance has been verified by comparison with existing studies.

The usage of Unmanned Aerial Vehicles (UAVs) has escalated widely in recent years. The mobility they provide and their capability to be used autonomously has made them applicable to a wide range of fields for both the civilians and the military. However, the flight safety is pivotal during UAV flights for any application and it is necessary that the recently ordained rules are complied with. Ensuring the compliance of such rules and regulations is especially difficult when using these UAVs are used in Autonomous mode. To establish the flight safety and compliance with the regulations, computer vision can be utilized while in autonomous mode. This is predominantly true when the flight is conducted in the proximity of crowds, animals, moving vehicles or when emergency and precision landings need to be performed. To that extent, we present a concept of drone flight control system which utilizes object detection and tracking algorithms.

객체 위치추적은 칼만필터나 루엔버거 추적기와 같은 다양한기법을 통해 연구되고 있는데, 시스템 모델이 명확하게 규정되지 않는경우와 같이 기존의 신호처리 기법을 성공적으로 적용하기 어려운 상황에서도 객체의 위치를 추적해 낼 수 있는 인공신경망을 설계하는 것이 가능하다. 본 논문에서는, 불규칙하게 운동 상태가 변화하는 객체의 위치를 지속적으로 추적하기 위해, 다양한 길이의 최대우도기법 객체위치 추정깂들을 도출한 뒤 신경망을 통해 이들을 적응적으로 가중평균하는 방법을 사용하는 '최대우도 가중평균 신경망'을 제안한다. 해당 신경망은 객체의 위치를 직접 추정하지 않고 데이터 길이가 다른 다양한 최대우도기법 추적 결과들을 가중평균하여 위치 추정을 수행한다. 우리는 제안하는 시스템의 추적성능을 칼만필터 및 최대우도기법들과 비교하여, 제안하는 기법이 물체의 움직임 특성에 잘 대처하여 우수한 성능을 나타내 줌을 보인다.

Object tracking is being studied with various techniques such as Kalman filter and Luenberger tracker. Even in situations, such as the one in which the system model is not well specified, to which existing signal processing techniques are not successfully applicable, it is possible to design artificial neural networks to track objects. In this paper, we propose an artificial neural network, which we call 'maximum-likelihood weighted-average neural network', to continuously track unpredictably moving objects. This neural network does not directly estimate the locations of an object but obtains location estimates by making weighted average combining various results of maximum likelihood tracking with different data lengths. We compare the performance of the proposed system with those of Kalman filter and maximum likelihood object trackers and show that the proposed scheme exhibits excellent performance well adapting the change of object moving characteristics.

객체추적은 이전시간에서 추정한 위치와 현재 관측 데이터를 바탕으로 객체의 위치를 연속적으로 추적하는 신호처리 분야이다. 이 논문에서는 3개의 RNN을 서브모듈로 가지는 적응형 스케일조절 신경망을 이용해 입력 데이터의 스케일을 스스로 조절하여 추적할 수 있는 신경망을 제안한다. 객체 추적 성능을 평가하기 위해 객체가 조각별 등가속운동을 하는 1차원 객체 운동 모델에서 제안하는 시스템, 칼만 필터와 최대우도기법의 추적 성능을 비교한다. 그 결과 제안하는 알고리듬의 성능이 평균제곱근오차 기준으로 최대우도기법과 칼만필터보다 다양한 상황에서 전반적으로 우수하며 관측잡음이 커질수록 성능격차가 더 커지는 것을 보인다.

Object tracking is a field of signal processing that sequentially tracks the location of an object based on the previous-time location estimations and the present-time observation data. In this paper, we propose an adaptive scaling neural network that can track and adjust the scale of the input data with three recursive neural network (RNN) submodules. To evaluate object tracking performance, we compare the proposed system with the Kalman filter and the maximum likelihood object tracking scheme under an one-dimensional object movement model in which the object moves with piecewise constant acceleration. We show that the proposed scheme is generally better, in terms of root mean square error (RMSE) performance, than maximum likelihood scheme and Kalman filter and that the performance gaps grow with increased observation noise.

본 논문에서는 컬러 히스토그램과 CNN 모델을 이용한 객체 추적 기법 알고리즘을 제안한다. CNN (convolutional neural network) 모델기반 객체 추적 알고리즘인 GOTURN (generic object tracking using regression network)의 정확도를 높이기 위해 컬러 히스토그램 기반 mean-shift 추적 알고리즘을 합성하였다. 두 알고리즘을 SVM (support vector machine)을 통해 분류하여 추적 정확도가 더 높은 알고리즘을 선택하도록 설계하였다. Mean-shift 추적 알고리즘은 객체 추적에 실패할 때 경계 박스가 큰 범위로 움직이는 경향이 있어 경계 박스의 이동거리에 제한을 두어 정확도를 향상시켰다. 또한 영상 평균 밝기, 히스토그램 유사도를 고려하여 두 알고리즘의 추적 시작 위치를 초기화하여 성능을 높였다. 결과적으로 기존 GOTURN 알고리즘보다 본 논문에서 제안한 알고리즘이 전체적으로 정확도가 1.6% 향상되었다.

In this paper, we propose an object tracking algorithm based on color histogram and convolutional neural network model. In order to increase the tracking accuracy, we synthesize generic object tracking using regression network algorithm which is one of the convolutional neural network model-based tracking algorithms and a mean-shift tracking algorithm which is a color histogram-based algorithm. Both algorithms are classified through support vector machine and designed to select an algorithm with higher tracking accuracy. The mean-shift tracking algorithm tends to move the bounding box to a large range when the object tracking fails, thus we improve the accuracy by limiting the movement distance of the bounding box. Also, we improve the performance by initializing the tracking start positions of the two algorithms based on the average brightness and the histogram similarity. As a result, the overall accuracy of the proposed algorithm is 1.6% better than the existing generic object tracking using regression network algorithm.

동영상에서의 물체추적은 각 프레임간의 시공간적 정보의 대응성을 추적함으로써 이루어질 수 있다. 시공간적 대응성을 결정함에 있어 각 프레임의 color invariant 속성을 이용하여 각 프레임의 물체 위치를 예측하고, 추출된 물체 블록의 누적 유사도 변환을 적용한 수치를 이용하여 프레임간의 물체 대응성을 결정함으로써 물체를 추적하는 방법론을 제시한다. 실험결과를 통하여 이러한 방법의 적용 적절성을 검증하였다.

In this paper, an object tracking system in a known environment is proposed. It extracts moving area shaped on objects in video sequences and decides tracks of moving objects. Color invarianoe features are exploited to extract the plausible object blocks and the degree of radial homogeneity, which is utilized as local block feature to find out the block correspondences. The experimental results are given.

객체추적은 객체의 위치변화를 찾는 것으로, 이전시간의 객체의 위치와 주어진 관측 데이터를 바탕으로 객체의 위치를 추적하는 신호처리의 한 분야이다. 객체추적 기법에는 대표적으로 칼만필터와 파티클필터가 있는데, 두 필터 모두 시스템 모델을 알고 있어야 좋은 성능을 낼 수 있다. 퍼셉트론 신경망에 피드백 루프를 추가한 재귀신경망은 데이터의 시계열적 상관관계를 활용할 수 있어 객체 추적에도 사용되고 있으며, 장기의존성 문제를 해결한 LSTM으로 발전하여 다양한 분야에 활용되고 있다. 본 논문에서는 이러한 LSTM의 추적 성능을 검증하기 위하여 1차원 객체 추적이라는 공통의 문제를 설정하고, 칼만필터, 파티클필터와의 추적 성능을 비교한다. 보다 다양한 관측 환경에서의 추적 성능 비교검증을 위하여 가우시안 잡음 외에도 라플라스, 지수, 균등 분포의 잡음이 있는 경우도 상정하였다. 그 결과 LSTM 신경망은 시스템 모델이 주어지지 않고, 학습데이터만으로 학습을 하여 안정적인 성능을 낼 수 있다는 것을 확인하였다.

Object tracking is a technique of signal processing that estimates objects locations based on past locations and present time observed data. While, Kalman filter and particle filter are among the most notable object tracking schemes, these filters need to know the system model to achieve optimal performance. The recursive neural network (RNN) with a feedback loop added to the perceptron neural network can be used for object tracking. Also, RNN evolved into long-short term memory (LSTM) that solved the long-term dependence problem and is being used in various fields. In this paper, in order to study the tracking performance of LSTM, we consider a simple problem of one-dimensional object tracking, and compare the tracking performance with Kalman and particle filters. In order to test the tracking performance in diverse observation environments, various noise models such as Gaussian, Laplace, exponential, and uniformly distributed noises are considered. Under the various circumstances, we observe that LSTM neural network achieves fairly stable performance without knowing the system model.

우리는 본 논문에서 이차 칼만 필터를 이용한 영상 기반 객체 추적분야의 새로운 알고리즘을 제안한다. 최근에 발표된 이차 칼만 필터는 영상 기반의 객체의 실제 3차원 공간의 위치를 추적하는 것에는 아직 적용되지 않았다. 2차원 영상 내의 위치를 3차원 공간상의 위치로 환원시키는 것은 비선형적 변환을 수반하기 때문에 그에 맞는 추적 알고리즘을 사용해야만 한다. 이러한 상황에서, 비선형 수식을 이차식으로 근사화하는 이차 칼만 필터가 선형으로 근사화하는 확장 칼만 필터보다 더 정확한 성능을 낼 수 있다. 우리는 동일한 상황을 가정하여 확장 칼만 필터, 무향 칼만 필터, 파티클 필터, 그리고 우리가 제안한 이차 칼만 필터를 이용하여 객체를 추적하고, 그 결과를 비교해 본다. 결론적으로 이차 칼만 필터가 발산율이 확장 칼만 필터에 비해 거의 절반가량 감소하며, 추적 정확도 측면에서 무향 칼만 필터에 비해 1% 가량 우수한 성능을 나타낸다.

In this paper, we propose a novel quadratic Kalman filter based object tracking algorithm using moving pictures. Quadratic Kalman filter, which is introduced recently, has not yet been applied to the problem of 3-dimensional (3-D) object tracking. Since the mapping of a position in 2-D moving pictures into a 3-D world involves non-linear transformation, appropriate algorithm must be chosen for object tracking. In this situation, the quadratic Kalman filter can achieve better accuracy than extended Kalman filter. Under the same conditions, we compare extended Kalman filter, unscented Kalman filter and sequential importance resampling particle filter together with the proposed scheme. In conculsion, the proposed scheme decreases the divergence rate by half compared with the scheme based on extended Kalman filter and improves the accuracy by about 1% in comparison with the one based on unscented Kalman filter.

실사 및 그래픽 기반 가상현실 콘텐츠는 360도 영상을 기반으로 하며, 시청자의 의도나 자동 추천 기능을 통한 뷰포트 추출이 필수적이다. 본 논문은 360도 영상에서 다중 객체 추적 기반의 뷰포트 추출 시스템을 설계하고, 다중 뷰포트 추출에 필요한 병렬화된 연산 구조를 제안한다. 360도 영상에서 뷰포트 추출 과정을 ERP 좌표의 3D 구 표면 좌표 변환과 3D 구 표면 좌표의 뷰포트 내 2D 좌표 변환 과정을 순서대로 픽셀 단위의 스레드로 구성하여 연산을 병렬화하였다. 제안 구조는 항공 360도 영상 시퀀스들에 대하여 최대 30개의 뷰포트 추출 과정에 대한 연산 시간이 평가되었으며, 뷰포트 수에 정비례하는 CPU 기반 연산 시간에 비해 최대 5240배 가속화됨을 확인하였다. ERP 프레임 I/O 시간을 줄일 수 있는 고속의 I/O나 메모리 버퍼를 사용 시 뷰포트 추출 시간을 7.82배 추가 가속화가 가능하다. 제안하는 뷰포트 추출 병렬화 구조는 360도 비디오나 가상현실 콘텐츠들에 대한 동시 다중 접속 서비스나 사용자별 영상 요약 서비스 등에 활용될 수 있다.

Realistic and graphics-based virtual reality content is based on 360-degree videos, and viewport extraction through the viewer's intention or automatic recommendation function is essential. This paper designs a viewport extraction system based on multiple object tracking in 360-degree videos and proposes a parallel computing structure necessary for multiple viewport extraction. The viewport extraction process in 360-degree videos is parallelized by composing pixel-wise threads, through 3D spherical surface coordinate transformation from ERP coordinates and 2D coordinate transformation of 3D spherical surface coordinates within the viewport. The proposed structure evaluated the computation time for up to 30 viewport extraction processes in aerial 360-degree video sequences and confirmed up to 5240 times acceleration compared to the CPU-based computation time proportional to the number of viewports. When using high-speed I/O or memory buffers that can reduce ERP frame I/O time, viewport extraction time can be further accelerated by 7.82 times. The proposed parallelized viewport extraction structure can be applied to simultaneous multi-access services for 360-degree videos or virtual reality contents and video summarization services for individual users.